Predicting Census Undercount Areas with Massive GPS Movement Pattern Data

Undercounting within the US National Census has always been a significant issue for specific populations, such as those with complex family structures. With the rise of Covid-19 and a shortened counting period, this year’s census is probably the most challenging to date. How can data science and new datasets help? We found that modern machine learning techniques can provide the basis for much-improved census undercount prediction. We also found that the addition of GPS data and point of interest data increased model accuracy and provided further insight into behavioral factors beyond demographics which significantly affect undercount.